platform/upstream/llvm.git
5 years agoRevert "[msan] Disable tail call optimization in msan rtl."
Evgeniy Stepanov [Thu, 27 Dec 2018 03:30:42 +0000 (03:30 +0000)]
Revert "[msan] Disable tail call optimization in msan rtl."

This reverts commit r350080, which breaks
* gcc as the host compiler
* some tests on ppc64

llvm-svn: 350083

5 years ago[X86] Factor the core code out of LowerSETCC into a helper that can create CMP/BT...
Craig Topper [Thu, 27 Dec 2018 01:50:40 +0000 (01:50 +0000)]
[X86] Factor the core code out of LowerSETCC into a helper that can create CMP/BT/PTEST/KORTEST etc. without making an X86ISD::SETCC node. NFCI

Make each of the helper functions only return their comparison node and the condition code. Leave X86ISD::SETCC creation to the LowerSETCC function itself.

Looking into whether we can use this code directly in BRCOND and SELECT lowering instead of going through LowerSETCC which creates an X86ISD::SETCC node we need to look through.

llvm-svn: 350082

5 years ago[X86] Merge getBitTestCondition into LowerAndToBT. Don't create X86ISD::SETCC node...
Craig Topper [Thu, 27 Dec 2018 01:50:38 +0000 (01:50 +0000)]
[X86] Merge getBitTestCondition into LowerAndToBT. Don't create X86ISD::SETCC node in the merged function. NFCI

Only one of the 3 callers of LowerAndToBT need the SETCC node. Two of them have to look through it to find the operands they really need. Instead create it after the one call that needs it.

LowerAndToBT now returns both the BT node and the X86 specific condition code separately.

llvm-svn: 350081

5 years ago[msan] Disable tail call optimization in msan rtl.
Evgeniy Stepanov [Thu, 27 Dec 2018 00:21:20 +0000 (00:21 +0000)]
[msan] Disable tail call optimization in msan rtl.

This should give us better error stack traces on the bots, in
particular, for https://bugs.llvm.org/show_bug.cgi?id=40162

llvm-svn: 350080

5 years ago[WebAssembly] Added basic support for if/else/end_if in MC layer.
Wouter van Oortmerssen [Wed, 26 Dec 2018 22:55:26 +0000 (22:55 +0000)]
[WebAssembly] Added basic support for if/else/end_if in MC layer.

Summary:
These instructions are currently unused in our backend, but for
completeness it is good to support them, so they can be used with
the assembler in hand-written code.

Tests are very basic, signature support missing much like other blocks.

Reviewers: dschuff, aheejin

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D55973

llvm-svn: 350079

5 years ago[WebAssembly] Make assembler check for proper nesting of control flow.
Wouter van Oortmerssen [Wed, 26 Dec 2018 22:46:18 +0000 (22:46 +0000)]
[WebAssembly] Make assembler check for proper nesting of control flow.

Summary:
It does so using a simple nesting stack, and gives clear errors upon
violation. This is unique to wasm, since most CPUs do not have
any nested constructs.

Had to add an end of file check to the general assembler for this.

Note: if/else/end instructions are not currently supported in our
tablegen defs, so these tests will be enabled in a follow-up.
They already pass the nesting check.

Reviewers: dschuff, aheejin

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D55797

llvm-svn: 350078

5 years agomanpages: Update the URL for https
Sylvestre Ledru [Wed, 26 Dec 2018 22:34:44 +0000 (22:34 +0000)]
manpages: Update the URL for https

llvm-svn: 350077

5 years ago[WebAssembly] Delete an unnecessary line in RegStackify
Heejin Ahn [Wed, 26 Dec 2018 22:33:35 +0000 (22:33 +0000)]
[WebAssembly] Delete an unnecessary line in RegStackify

`OneUseInst` is set outside of the loop before and `OneUse` does not
change throughout the loop, so this line is not necessary.

llvm-svn: 350076

5 years ago[WebAssembly] Fix typos in comments in RegStackify (NFC)
Heejin Ahn [Wed, 26 Dec 2018 22:27:46 +0000 (22:27 +0000)]
[WebAssembly] Fix typos in comments in RegStackify (NFC)

llvm-svn: 350075

5 years ago[LoopIdiomRecognize] Add CTTZ support
Craig Topper [Wed, 26 Dec 2018 21:59:48 +0000 (21:59 +0000)]
[LoopIdiomRecognize] Add CTTZ support

Summary:
Existing LIR recognizes CTLZ where shifting input variable right until it is zero. (Shift-Until-Zero idiom)

This commit:
1. Augments Shift-Until-Zero idiom to recognize CTTZ where input variable is shifted left.
2. Prepare for BitScan idiom recognition.

Patch by Yuanfang Chen (tabloid.adroit)

Reviewers: craig.topper, evstupac

Reviewed By: craig.topper

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D55876

llvm-svn: 350074

5 years ago[codeview] Check if this 'this' type of a method is a pointer
Reid Kleckner [Wed, 26 Dec 2018 21:52:17 +0000 (21:52 +0000)]
[codeview] Check if this 'this' type of a method is a pointer

Fixes crash reported after r347354 for frontends that don't always emit
'this' pointers for methods. Now we will silently produce debug info
that makes functions like this look like static methods, which seems
reasonable.

llvm-svn: 350073

5 years ago[clang-cl] Treat inputs as C++ with /E, like MSVC
Reid Kleckner [Wed, 26 Dec 2018 21:04:08 +0000 (21:04 +0000)]
[clang-cl] Treat inputs as C++ with /E, like MSVC

midl invokes the compiler on .idl files with /E. Before this change, we
would treat unrecognized inputs as object files. Now we pre-process to
stdout as expected. I checked that MSVC defines __cplusplus when invoked
this way, so treating the input as C++ seems like the right thing to do.

After this change, I was able to run midl like this with clang-cl:
$ midl -cpp_cmd clang-cl.exe foo.idl

Things worked for the example IDL file in the Microsoft documentation,
but beyond that, I don't know if this will work well.

Fixes PR40140

llvm-svn: 350072

5 years ago[MS] Mangle return adjusting thunks with the public access specifier
Reid Kleckner [Wed, 26 Dec 2018 20:07:52 +0000 (20:07 +0000)]
[MS] Mangle return adjusting thunks with the public access specifier

MSVC does this, so we should too.

Fixes PR40138

llvm-svn: 350071

5 years agoUse error() instead of fatal() to report an invalid address range.
Rui Ueyama [Wed, 26 Dec 2018 19:15:04 +0000 (19:15 +0000)]
Use error() instead of fatal() to report an invalid address range.

In this patch we also use toString() to stringize a section.

llvm-svn: 350070

5 years ago[NVPTX] Allow libcalls that are defined in the current module.
Justin Lebar [Wed, 26 Dec 2018 19:12:31 +0000 (19:12 +0000)]
[NVPTX] Allow libcalls that are defined in the current module.

The patch adds a possibility to make library calls on NVPTX.

An important thing about library functions - they must be defined within
the current module. This basically should guarantee that we produce a
valid PTX assembly (without calls to not defined functions). The one who
wants to use the libcalls is probably will have to link against
compiler-rt or any other implementation.

Currently, it's completely impossible to make library calls because of
error LLVM ERROR: Cannot select: i32 = ExternalSymbol '...'. But we can
lower ExternalSymbol to TargetExternalSymbol and verify if the function
definition is available.

Also, there was an issue with a DAG during legalisation. When we expand
instruction into libcall, the inner call-chain isn't being "integrated"
into outer chain. Since the last "data-flow" (call retval load) node is
located in call-chain earlier than CALLSEQ_END node, the latter becomes
a leaf and therefore a dead node (and is being removed quite fast).
Proposed here solution relies on another data-flow pseudo nodes
(ProxyReg) which purpose is only to keep CALLSEQ_END at legalisation and
instruction selection phases - we remove the pseudo instructions before
register scheduling phase.

Patch by Denys Zariaiev!

Differential Revision: https://reviews.llvm.org/D34708

llvm-svn: 350069

5 years agoIgnore ConstantExpr in IgnoreParens
Reid Kleckner [Wed, 26 Dec 2018 17:44:40 +0000 (17:44 +0000)]
Ignore ConstantExpr in IgnoreParens

Summary:
This moves it up from IgnoreParenImpCasts to IgnoreParens, so that more
helpers ignore it. For most clients, this ensures that these helpers
behave the same with and without C++17 enabled, which is what appears to
introduce these new expression nodes.

Fixes PR39881

Reviewers: void, rsmith

Subscribers: cfe-commits

Differential Revision: https://reviews.llvm.org/D55853

llvm-svn: 350068

5 years agoPass a concrete triple for two OpenMP tests that depend on TLS
Nico Weber [Wed, 26 Dec 2018 16:06:26 +0000 (16:06 +0000)]
Pass a concrete triple for two OpenMP tests that depend on TLS

Not all %itanium_abi_triple values support TLS. Makes
OpenMP/declare_reduction_codegen.cpp, OpenMP/parallel_copyin_codegen.cpp for
%itanium_abi_triples without TLS support.

Alternatively we could pass -fnoopenmp-use-tls and tweak some of the CHECK
lines, but this seems a bit simpler.

Fixes PR40156.

Differential Revision: https://reviews.llvm.org/D56086

llvm-svn: 350067

5 years ago[AMDGPU] Regenerate i64 shift tests.
Simon Pilgrim [Wed, 26 Dec 2018 12:09:10 +0000 (12:09 +0000)]
[AMDGPU] Regenerate i64 shift tests.

To show codegen diff due to a future SimplifyDemandedBits patch.

llvm-svn: 350065

5 years ago[NFC] Use utility function for guards detection
Max Kazantsev [Wed, 26 Dec 2018 08:22:25 +0000 (08:22 +0000)]
[NFC] Use utility function for guards detection

llvm-svn: 350064

5 years ago[MIPS GlobalISel] Select G_SELECT
Petar Avramovic [Tue, 25 Dec 2018 14:42:30 +0000 (14:42 +0000)]
[MIPS GlobalISel] Select G_SELECT

Add widen scalar for type index 1 (i1 condition) for G_SELECT.
Select G_SELECT for pointer, s32(integer) and smaller low level
types on MIPS32.

Differential Revision: https://reviews.llvm.org/D56001

llvm-svn: 350063

5 years ago[NFC] Reuse variables instead of re-calling getParent
Max Kazantsev [Tue, 25 Dec 2018 07:20:06 +0000 (07:20 +0000)]
[NFC] Reuse variables instead of re-calling getParent

llvm-svn: 350062

5 years ago[PowerPC] Fix the bug of ISD::ADDE to set its second return type to glue
Kang Zhang [Tue, 25 Dec 2018 03:29:51 +0000 (03:29 +0000)]
[PowerPC] Fix the bug of ISD::ADDE to set its second return type to glue

Summary:
This patch is to fix the bug imported by rL341634.
In above submit , the the return type of ISD::ADDE is
14224: SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i64),
but in fact, the second return type of ISD::ADDE should be
MVT::Glue not MVT::i64.

Reviewed By: hfinkel

Differential Revision: https://reviews.llvm.org/D55977

llvm-svn: 350061

5 years ago[gn build] Make NOSORT line actually work
Nico Weber [Mon, 24 Dec 2018 23:06:29 +0000 (23:06 +0000)]
[gn build] Make NOSORT line actually work

GN wants the NOSORT line to be the first line of a comment block, not the last
line.

I sent https://gn-review.googlesource.com/c/gn/+/3560 to support having it in
the last line too, but since it will be a while until everyone has that change
even if it's expected, use the form that works today.

Differential Revision: https://reviews.llvm.org/D56065

llvm-svn: 350060

5 years ago[X86] Use GetDemandedBits to simplify the operands of PMULDQ/PMULUDQ.
Craig Topper [Mon, 24 Dec 2018 19:40:20 +0000 (19:40 +0000)]
[X86] Use GetDemandedBits to simplify the operands of PMULDQ/PMULUDQ.

This is an alternative to what I attempted in D56057.

GetDemandedBits is a special version of SimplifyDemandedBits that allows simplifications even when the operand has other uses. GetDemandedBits will only do simplifications that allow a node to be bypassed. It won't create new nodes or alter any of the other users.

I had to add support for bypassing SIGN_EXTEND_INREG to GetDemandedBits.

Based on a patch that Simon Pilgrim sent me in email.

Fixes PR40142.

llvm-svn: 350059

5 years ago[X86] Add test cases for PR40142. NFC
Craig Topper [Mon, 24 Dec 2018 19:40:17 +0000 (19:40 +0000)]
[X86] Add test cases for PR40142. NFC

llvm-svn: 350058

5 years ago[analyzer] [NFC] Clean up the mess of constructing argument effects in RetainCountChecker
George Karpenkov [Mon, 24 Dec 2018 18:45:18 +0000 (18:45 +0000)]
[analyzer] [NFC] Clean up the mess of constructing argument effects in RetainCountChecker

Previously, argument effects were stored in a method variable, which was
effectively global.
The global state was reset at each (hopefully) entrance point to the
summary construction,
and every function could modify it.

Differential Revision: https://reviews.llvm.org/D56036

llvm-svn: 350057

5 years ago[clang-tidy] add IgnoreMacros option to readability-uppercase-literal-suffix
Miklos Vajna [Mon, 24 Dec 2018 17:47:32 +0000 (17:47 +0000)]
[clang-tidy] add IgnoreMacros option to readability-uppercase-literal-suffix

And also enable it by default to be consistent with e.g.
modernize-use-using.

This helps e.g. when running this check on client code where the macro
is provided by the system, so there is no easy way to modify it.

Reviewed By: JonasToth, lebedev.ri

Differential Revision: https://reviews.llvm.org/D56025

llvm-svn: 350056

5 years ago[HWASAN] Instrument memorty intrinsics by default
Eugene Leviant [Mon, 24 Dec 2018 16:02:48 +0000 (16:02 +0000)]
[HWASAN] Instrument memorty intrinsics by default

Differential revision: https://reviews.llvm.org/D55926

llvm-svn: 350055

5 years ago[gn build] Add build files for clang/tools/{c-arcmt-test,c-index-test} and their...
Nico Weber [Mon, 24 Dec 2018 15:45:04 +0000 (15:45 +0000)]
[gn build] Add build files for clang/tools/{c-arcmt-test,c-index-test} and their dependency clang/tools/libclang

libclang is somewhat incomplete. It's just enough to get check-clang to pass,
but that requires it to be pretty complete. The biggest thing is that it's not
built as a shared library on Linux. The libclang/BUILD.gn file has a comment
with details on what else is missing.

Differential Revision: https://reviews.llvm.org/D56059

llvm-svn: 350054

5 years agoReleaseNotes: X86 Target: bdver2 sched model was added (D52779)
Roman Lebedev [Mon, 24 Dec 2018 12:12:26 +0000 (12:12 +0000)]
ReleaseNotes: X86 Target: bdver2 sched model was added (D52779)

llvm-svn: 350053

5 years agoRevert rL350048 and rL350050
Max Kazantsev [Mon, 24 Dec 2018 10:30:04 +0000 (10:30 +0000)]
Revert rL350048 and rL350050

These patches have broken almost all buildbots on test
DebugInfo/X86/addr_comments.ll. Reverting to green.

llvm-svn: 350052

5 years ago[LLD][ELF] - Cleanup gdb-index-dwarf5-low-high.s. NFC.
George Rimar [Mon, 24 Dec 2018 10:18:07 +0000 (10:18 +0000)]
[LLD][ELF] - Cleanup gdb-index-dwarf5-low-high.s. NFC.

Remove tabs and .debug_macinfo sections.

llvm-svn: 350051

5 years agoFix build - follow-up to r350048 which broke headerless (v4) address pool
David Blaikie [Mon, 24 Dec 2018 07:56:40 +0000 (07:56 +0000)]
Fix build - follow-up to r350048 which broke headerless (v4) address pool

llvm-svn: 350050

5 years ago[LoopSimplifyCFG] Delete dead exiting edges
Max Kazantsev [Mon, 24 Dec 2018 07:41:33 +0000 (07:41 +0000)]
[LoopSimplifyCFG] Delete dead exiting edges

This patch teaches LoopSimplifyCFG to remove dead exiting edges
from loops.

Differential Revision: https://reviews.llvm.org/D54025
Reviewed By: fedor.sergeev

llvm-svn: 350049

5 years agoDebugInfo: Use assembly label arithmetic for address pool size for easier reading...
David Blaikie [Mon, 24 Dec 2018 07:35:10 +0000 (07:35 +0000)]
DebugInfo: Use assembly label arithmetic for address pool size for easier reading/editing

llvm-svn: 350048

5 years agoDebugInfo: Add assembly comments for debug_addr contribution header fields
David Blaikie [Mon, 24 Dec 2018 07:09:50 +0000 (07:09 +0000)]
DebugInfo: Add assembly comments for debug_addr contribution header fields

llvm-svn: 350047

5 years agollvm-dwarfdump: Skip address index info (and dump only the address, if found) when...
David Blaikie [Mon, 24 Dec 2018 06:52:31 +0000 (06:52 +0000)]
llvm-dwarfdump: Skip address index info (and dump only the address, if found) when non-verbose dumping addrx forms

There's a few bugs here still - demonstrated with FIXITs in the test.

llvm-svn: 350046

5 years agoReturn "[LoopSimplifyCFG] Delete dead in-loop blocks"
Max Kazantsev [Mon, 24 Dec 2018 06:06:17 +0000 (06:06 +0000)]
Return "[LoopSimplifyCFG] Delete dead in-loop blocks"

The underlying bug that caused the revert should be fixed by rL348567.

Differential Revision: https://reviews.llvm.org/D54023

llvm-svn: 350045

5 years ago[LoopIdioms] More LocationSize::precise annotations; NFC
George Burgess IV [Mon, 24 Dec 2018 05:55:50 +0000 (05:55 +0000)]
[LoopIdioms] More LocationSize::precise annotations; NFC

Both of these places reference memset-like loops. Memset is precise.

Trying to keep these patches super small so they're easily post-commit
verifiable, as requested in D44748.

llvm-svn: 350044

5 years ago[X86] Remove unused variables left after r350041. NFC
Craig Topper [Mon, 24 Dec 2018 05:45:45 +0000 (05:45 +0000)]
[X86] Remove unused variables left after r350041. NFC

llvm-svn: 350043

5 years ago[SelectionDAGBuilder] Use ::precise LocationSizes; NFC
George Burgess IV [Mon, 24 Dec 2018 05:34:21 +0000 (05:34 +0000)]
[SelectionDAGBuilder] Use ::precise LocationSizes; NFC

More migration so we can disable the implicit int -> LocationSize
conversion.

All of these are either scatter/gather'ed vector instructions, or direct
loads. Hence, they're all precise.

Perhaps if we see way more getTypeStoreSize calls, we can make a
getTypeStoreLocationSize (or similar) as a wrapper that applies this
::precise. Doesn't appear that it's a good idea to make getTypeStoreSize
return a LocationSize itself, however.

llvm-svn: 350042

5 years ago[X86] Move the optimization that turns 'CMP (AND+IMM64), 0' into SRL/SHL+TEST to...
Craig Topper [Mon, 24 Dec 2018 05:27:13 +0000 (05:27 +0000)]
[X86] Move the optimization that turns 'CMP (AND+IMM64), 0' into SRL/SHL+TEST to X86ISelDAGToDAG.

This cleans more code out of EmitTest.

llvm-svn: 350041

5 years ago[clangd] Delete stray semicolon. NFC
Fangrui Song [Mon, 24 Dec 2018 03:19:53 +0000 (03:19 +0000)]
[clangd] Delete stray semicolon. NFC

llvm-svn: 350040

5 years ago[X86] Autogenerate complete checks. NFC
Craig Topper [Mon, 24 Dec 2018 01:59:31 +0000 (01:59 +0000)]
[X86] Autogenerate complete checks. NFC

llvm-svn: 350039

5 years ago[X86] Remove the ANDN check from EmitTest.
Craig Topper [Mon, 24 Dec 2018 01:10:13 +0000 (01:10 +0000)]
[X86] Remove the ANDN check from EmitTest.

Remove the TESTmr isel patterns and add another postprocessing combine for TESTrr+ANDrm->TESTmr. We already have a postprocessing combine for TESTrr+ANDrr->TESTrr. With this we can give ANDN a chance to match first. And clean it up during post processing if we ended up with just a regular AND.

This is another step towards my plan to gut EmitTest and do more flag handling during isel matching or by using optimizeCompare.

llvm-svn: 350038

5 years ago[clangd] Delete trailing ; NFC
Fangrui Song [Sun, 23 Dec 2018 22:20:34 +0000 (22:20 +0000)]
[clangd] Delete trailing ; NFC

llvm-svn: 350037

5 years agoTolerate flaky LLD tests on NetBSD
Kamil Rytarowski [Sun, 23 Dec 2018 21:20:09 +0000 (21:20 +0000)]
Tolerate flaky LLD tests on NetBSD

Summary:
For unknown reasons  LLD tests are flaky on the NetBSD buildbot,
but not on local machines of developers.

Unless the linker will be fully functional on this target,
allow to pass flaky tests with optional retry.

Reviewers: joerg, mgorny, ruiu

Reviewed By: mgorny

Subscribers: lebedev.ri, MaskRay, llvm-commits, #lld

Tags: #lld

Differential Revision: https://reviews.llvm.org/D56053

llvm-svn: 350036

5 years ago[llvm-exegesis] Clustering: don't enqueue a point multiple times
Fangrui Song [Sun, 23 Dec 2018 20:48:52 +0000 (20:48 +0000)]
[llvm-exegesis] Clustering: don't enqueue a point multiple times

Summary:
SetVector uses both DenseSet and vector, which is time/memory inefficient. The points are represented as natural numbers so we can replace the DenseSet part by indexing into a vector<char> instead.

Don't cargo cult the pseudocode on the wikipedia DBSCAN page. This is a standard BFS style algorithm (the similar loops have been used several times in other LLVM components): every point is processed at most once, thus the queue has at most NumPoints elements. We represent it with a vector and allocate it outside of the loop to avoid allocation in the loop body.

We check `Processed[P]` to avoid enqueueing a point more than once, which also nicely saves us a `ClusterIdForPoint_[Q].isUndef()` check.

Many people hate the oneshot abstraction but some favor it, therefore we make a compromise, use a lambda to abstract away the neighbor adding process.

Delete the comment `assert(Neighbors.capacity() == (Points_.size() - 1));` as it is wrong.

llvm-svn: 350035

5 years ago[DAGCombiner] limit shuffle to extend transform (PR40146)
Sanjay Patel [Sun, 23 Dec 2018 20:48:31 +0000 (20:48 +0000)]
[DAGCombiner] limit shuffle to extend transform (PR40146)

It's dangerous to knowingly create an illegal vector type
no matter what stage of combining we're in.

This prevents the missed folding/scalarization seen in:
https://bugs.llvm.org/show_bug.cgi?id=40146

llvm-svn: 350034

5 years ago[x86] add test for vector shuffle --> extend transform (PR40146); NFC
Sanjay Patel [Sun, 23 Dec 2018 20:36:52 +0000 (20:36 +0000)]
[x86] add test for vector shuffle --> extend transform (PR40146); NFC

llvm-svn: 350033

5 years ago[DAGCombiner] allow hoisting vector bitwise logic ahead of extends
Sanjay Patel [Sun, 23 Dec 2018 19:58:16 +0000 (19:58 +0000)]
[DAGCombiner] allow hoisting vector bitwise logic ahead of extends

llvm-svn: 350032

5 years ago[x86] add tests for vector extend + logic ops; NFC
Sanjay Patel [Sun, 23 Dec 2018 18:37:44 +0000 (18:37 +0000)]
[x86] add tests for vector extend + logic ops; NFC

llvm-svn: 350031

5 years ago[xray] Detect MPROTECT and error out when it's enabled (on NetBSD)
Michal Gorny [Sun, 23 Dec 2018 15:09:28 +0000 (15:09 +0000)]
[xray] Detect MPROTECT and error out when it's enabled (on NetBSD)

Add a CheckMPROTECT() routine to detect when pax MPROTECT is enabled
on NetBSD, and error xray out when it is.  The solution is adapted
from existing CheckASLR().

Differential Revision: https://reviews.llvm.org/D56049

llvm-svn: 350030

5 years ago[xray] Disable alignas() for thread_local objects on NetBSD
Michal Gorny [Sun, 23 Dec 2018 15:09:20 +0000 (15:09 +0000)]
[xray] Disable alignas() for thread_local objects on NetBSD

Disable enforcing alignas() for structs that are used as thread_local
data on NetBSD.  The NetBSD ld.so implementation is buggy and does
not enforce correct alignment; however, clang seems to take it for
granted and generates instructions that segv on wrongly aligned objects.
Therefore, disable those alignas() statements on NetBSD until we can
establish a better fix.

Apparently, std::aligned_storage<> does not have any real effect
at the moment, so we can leave it as-is.

Differential Revision: https://reviews.llvm.org/D56000

llvm-svn: 350029

5 years ago[Driver] Disable -faddrsig on Gentoo by default
Michal Gorny [Sun, 23 Dec 2018 15:07:26 +0000 (15:07 +0000)]
[Driver] Disable -faddrsig on Gentoo by default

Gentoo supports combining clang toolchain with GNU binutils, and many
users actually do that.  As -faddrsig is not supported by GNU strip,
this results in a lot of warnings.  Disable it by default and let users
enable it explicitly if they want it; with the intent of reevaluating
when the underlying feature becomes standarized.

See also: https://bugs.gentoo.org/667854

Differential Revision: https://reviews.llvm.org/D56047

llvm-svn: 350028

5 years ago[Distro] Support detecting Gentoo
Michal Gorny [Sun, 23 Dec 2018 15:07:19 +0000 (15:07 +0000)]
[Distro] Support detecting Gentoo

Add support for distinguishing plain Gentoo distribution, and a unit
test for it.  This is going to be used to introduce distro-specific
customizations in the driver code; most notably, it is going to be used
to disable -faddrsig.

Differential Revision: https://reviews.llvm.org/D56024

llvm-svn: 350027

5 years ago[gn build] Add build files for clang/tools/{arcmt-test,clang-check,clang-func-mapping}
Nico Weber [Sun, 23 Dec 2018 14:19:53 +0000 (14:19 +0000)]
[gn build] Add build files for clang/tools/{arcmt-test,clang-check,clang-func-mapping}

Needed for check-clang.

Differential Revision: https://reviews.llvm.org/D56056

llvm-svn: 350026

5 years ago[gn build] Add build files for clang/tools/{clang-refactor,clang-rename}, clang/utils...
Nico Weber [Sun, 23 Dec 2018 14:17:13 +0000 (14:17 +0000)]
[gn build] Add build files for clang/tools/{clang-refactor,clang-rename}, clang/utils/hmaptool, clang/lib/Tooling/Refactoring

Needed for check-clang.

Differential Revision: https://reviews.llvm.org/D56055

llvm-svn: 350025

5 years ago[gn build] Add build files for clang/tools/{clang-diff,clang-import-test,diagtool...
Nico Weber [Sun, 23 Dec 2018 14:15:26 +0000 (14:15 +0000)]
[gn build] Add build files for clang/tools/{clang-diff,clang-import-test,diagtool and clang/lib/Tooling, clang/lib/Tooling/ASTDiff

Needed for check-clang.

Differential Revision: https://reviews.llvm.org/D56054

llvm-svn: 350024

5 years ago[ORC] Rename register in the OrcMips64 resolver code comments. NFC
Simon Atanasyan [Sun, 23 Dec 2018 12:05:04 +0000 (12:05 +0000)]
[ORC] Rename register in the OrcMips64 resolver code comments. NFC

The `fp` and `s8` register names are synonyms. But `fp` better reflects
a purpose of the register.

llvm-svn: 350023

5 years ago[ORC] clang-format OrcMips32 and OrcMips64 code. NFC
Simon Atanasyan [Sun, 23 Dec 2018 12:05:00 +0000 (12:05 +0000)]
[ORC] clang-format OrcMips32 and OrcMips64 code. NFC

llvm-svn: 350022

5 years ago[ORC] Remove redundant instruction from MIPS resolver code. NFC
Simon Atanasyan [Sun, 23 Dec 2018 12:04:55 +0000 (12:04 +0000)]
[ORC] Remove redundant instruction from MIPS resolver code. NFC

It's redundant to restore the `$a3` register twice.

llvm-svn: 350021

5 years agoSet the default SANITIZER_CXX_ABI library to libc++ for FreeBSD, and
Dimitry Andric [Sun, 23 Dec 2018 11:49:47 +0000 (11:49 +0000)]
Set the default SANITIZER_CXX_ABI library to libc++ for FreeBSD, and
also mark it as a system library, like on macOS.

llvm-svn: 350020

5 years ago[MemCpyOpt] Use LocationSize instead of ints; NFC
George Burgess IV [Sun, 23 Dec 2018 06:40:39 +0000 (06:40 +0000)]
[MemCpyOpt] Use LocationSize instead of ints; NFC

Trying to keep these patches super small so they're easily post-commit
verifiable, as requested in D44748.

srcSize is derived from the size of an alloca, and we quit out if the
size of that is > the size of the thing we're copying to. Hence, we
should always copy everything over, so these sizes are precise.

Don't make srcSize itself a LocationSize, since optionality isn't
helpful, and we do some comparisons against other sizes elsewhere in
that function.

llvm-svn: 350019

5 years ago[X86] Return false from hasAndNotCompare if the comparision value is a constant.
Craig Topper [Sun, 23 Dec 2018 05:52:55 +0000 (05:52 +0000)]
[X86] Return false from hasAndNotCompare if the comparision value is a constant.

We won't end up using an ANDN instruction in this case so we should generate the same code we do for pre-BMI targets.

llvm-svn: 350018

5 years ago[MemoryLocation] Use LocationSize instead of ints; NFC
George Burgess IV [Sun, 23 Dec 2018 03:36:44 +0000 (03:36 +0000)]
[MemoryLocation] Use LocationSize instead of ints; NFC

Trying to keep these patches super small so they're easily post-commit
verifiable, as requested in D44748.

This one sadly isn't *super* small, but all of the changes here are
either to:
- libfuncs that are passed a constant size (memcpy, memset, ...)
- instructions that store/load a constant size

So they have to be precise

llvm-svn: 350017

5 years ago[Loads] Use LocationSize instead of ints; NFC
George Burgess IV [Sun, 23 Dec 2018 03:10:56 +0000 (03:10 +0000)]
[Loads] Use LocationSize instead of ints; NFC

Keeping these patches super small so they're easily post-commit
verifiable, as requested in D44748.

This tries to find literal loads/stores of the given type, so this has
to be precise.

llvm-svn: 350016

5 years ago[Lint] Use LocationSize instead of ints; NFC
George Burgess IV [Sun, 23 Dec 2018 02:50:08 +0000 (02:50 +0000)]
[Lint] Use LocationSize instead of ints; NFC

Keeping these patches super small so they're easily post-commit
verifiable, as requested in D44748.

llvm-svn: 350015

5 years ago[AAEval] Use LocationSize instead of ints; NFC
George Burgess IV [Sun, 23 Dec 2018 02:39:58 +0000 (02:39 +0000)]
[AAEval] Use LocationSize instead of ints; NFC

Keeping these patches super small so they're easily post-commit
verifiable, as requested in D44748.

llvm-svn: 350014

5 years ago[X86] Fix an old FIXME about folding the zero constant into the OR instruction we...
Craig Topper [Sun, 23 Dec 2018 01:54:43 +0000 (01:54 +0000)]
[X86] Fix an old FIXME about folding the zero constant into the OR instruction we use for sequentially consistent fence in 32-bit mode without SSE2.

llvm-svn: 350013

5 years ago[X86] Autogenerate complete checks. NFC
Craig Topper [Sun, 23 Dec 2018 01:54:41 +0000 (01:54 +0000)]
[X86] Autogenerate complete checks. NFC

llvm-svn: 350012

5 years agoTest DWARFv5 with gdb-index and low_pc/high_pc on the CU (rather than ranges)
David Blaikie [Sat, 22 Dec 2018 22:20:47 +0000 (22:20 +0000)]
Test DWARFv5 with gdb-index and low_pc/high_pc on the CU (rather than ranges)

There was a bug in LLVM's libDebugInfo where it did not porpagate the
section index through the range query built from low_pc/high_pc. Hard to
test in LLVM, so I'm adding a test here.

llvm-svn: 350011

5 years agoDebugInfo: Accurately propagate the section used by a relocation when accessing range...
David Blaikie [Sat, 22 Dec 2018 22:20:40 +0000 (22:20 +0000)]
DebugInfo: Accurately propagate the section used by a relocation when accessing ranges defined by low/high_pc

This is difficult/not possible to test in LLVM, but is visible as a
crash in LLD when parsing DWARF to generate gdb-index.

This function is called by llvm-dwarfdump when parsing high_pc for
non-verbose output (to print the actual high_pc rather than the low_pc
relative value), but in that case llvm-dwarfdump doesn't print section
names (if it did, it would hit this problem).

We could add some other features to llvm-dwarfdump to expose this, but
nothing really springs to my mind. I will add a test to lld, though.

llvm-svn: 350010

5 years agollvm-dwarfdump: Dump the section name/number for addr attributes
David Blaikie [Sat, 22 Dec 2018 20:34:58 +0000 (20:34 +0000)]
llvm-dwarfdump: Dump the section name/number for addr attributes

llvm-svn: 350009

5 years ago[Analysis] More LocationSize cleanup; NFC
George Burgess IV [Sat, 22 Dec 2018 18:23:21 +0000 (18:23 +0000)]
[Analysis] More LocationSize cleanup; NFC

Keeping these patches super small so they're easily post-commit
verifiable, as requested in D44748.

llvm-svn: 350008

5 years ago[Analysis] s/uint64_t/LocationSize; NFC
George Burgess IV [Sat, 22 Dec 2018 17:42:08 +0000 (17:42 +0000)]
[Analysis] s/uint64_t/LocationSize; NFC

Let the gradual cleanup from D44748 continue. :)

llvm-svn: 350007

5 years ago[DAGCombiner] allow narrowing of add followed by truncate
Sanjay Patel [Sat, 22 Dec 2018 17:10:31 +0000 (17:10 +0000)]
[DAGCombiner] allow narrowing of add followed by truncate

trunc (add X, C ) --> add (trunc X), C'

If we're throwing away the top bits of an 'add' instruction, do it in the narrow destination type.
This makes the truncate-able opcode list identical to the sibling transform done in IR (in instcombine).

This change used to show regressions for x86, but those are gone after D55494.
This gets us closer to deleting the x86 custom function (combineTruncatedArithmetic)
that does almost the same thing.

Differential Revision: https://reviews.llvm.org/D55866

llvm-svn: 350006

5 years ago[x86] add load fold patterns for movddup with vzext_load
Sanjay Patel [Sat, 22 Dec 2018 16:59:02 +0000 (16:59 +0000)]
[x86] add load fold patterns for movddup with vzext_load

The missed load folding noticed in D55898 is visible independent of that change
either with an adjusted IR pattern to start or with AVX2/AVX512 (where the build
vector becomes a broadcast first; movddup is not produced until we get into isel
via tablegen patterns).

Differential Revision: https://reviews.llvm.org/D55936

llvm-svn: 350005

5 years agoImproving this fatal diagnostic to help checker developers figure out what's actually...
Aaron Ballman [Sat, 22 Dec 2018 15:31:57 +0000 (15:31 +0000)]
Improving this fatal diagnostic to help checker developers figure out what's actually gone wrong when they hit it.

llvm-svn: 350004

5 years ago[AST] Store the arguments of CXXConstructExpr in a trailing array
Bruno Ricci [Sat, 22 Dec 2018 14:39:30 +0000 (14:39 +0000)]
[AST] Store the arguments of CXXConstructExpr in a trailing array

Store the arguments of CXXConstructExpr in a trailing array. This is very
similar to the CallExpr case in D55771, with the exception that there is
only one derived class (CXXTemporaryObjectExpr) and that we compute the
offset to the trailing array instead of storing it.

This saves one pointer per CXXConstructExpr and CXXTemporaryObjectExpr.

Reviewed By: rjmccall

Differential Revision: https://reviews.llvm.org/D56022

llvm-svn: 350003

5 years ago[Sanitizer] Enable POSIX regex api on FreeBSD.
David Carlier [Sat, 22 Dec 2018 11:17:27 +0000 (11:17 +0000)]
[Sanitizer] Enable POSIX regex api on FreeBSD.

Reviewers: krytarowski

Reviewed By: krytarowski

Differential Revision: https://reviews.llvm.org/D56009

M    lib/sanitizer_common/sanitizer_common_interceptors.inc
M    lib/sanitizer_common/sanitizer_platform_interceptors.h
M    lib/sanitizer_common/sanitizer_platform_limits_freebsd.cc
M    lib/sanitizer_common/sanitizer_platform_limits_freebsd.h
D    test/sanitizer_common/TestCases/NetBSD/regex.cc
A  + test/sanitizer_common/TestCases/Posix/regex.cc

llvm-svn: 350002

5 years ago[runtime] [test] Fix using %python path
Michal Gorny [Sat, 22 Dec 2018 10:51:53 +0000 (10:51 +0000)]
[runtime] [test] Fix using %python path

Fix the newly-added tests to use %python substitution in order to use
the correct path to Python interpreter.  Otherwise, they fail on NetBSD
where there is no 'python', just 'pythonX.Y'.

Differential Revision: https://reviews.llvm.org/D56048

llvm-svn: 350001

5 years agoNFC][CodeGen][X86][AArch64] Tests for bit extract (pat. a/c/d) with trunc (PR36419)
Roman Lebedev [Sat, 22 Dec 2018 10:38:05 +0000 (10:38 +0000)]
NFC][CodeGen][X86][AArch64] Tests for bit extract (pat. a/c/d) with trunc (PR36419)

llvm-svn: 350000

5 years ago[NFC][CodeGen][X86][AArch64] Bit extract: add nounwind attr to drop .cfi noise
Roman Lebedev [Sat, 22 Dec 2018 09:58:13 +0000 (09:58 +0000)]
[NFC][CodeGen][X86][AArch64] Bit extract: add nounwind attr to drop .cfi noise

Forgot about that.

llvm-svn: 349999

5 years ago[NFC][CodeGen][X86][AArch64] Tests for bit extract (pat. b) with trunc (PR36419)
Roman Lebedev [Sat, 22 Dec 2018 09:40:14 +0000 (09:40 +0000)]
[NFC][CodeGen][X86][AArch64] Tests for bit extract (pat. b) with trunc (PR36419)

@bextr64_32_b1 is extracted from hotpath of real-world code
(RawSpeed BitStream<>::peekBitsNoFill()) after `clang -O3`.

@bextr64_32_b2/@bextr64_32_b0 is the same pattern,
but with trunc done last, showing how i think it can be handled:
https://rise4fun.com/Alive/K4B
https://rise4fun.com/Alive/qC9

It is possible that middle-end should do some of this, too.

https://bugs.llvm.org/show_bug.cgi?id=36419

llvm-svn: 349998

5 years agollvm-dwarfdump: Remove extraneous space between '(' and 'indexed'
David Blaikie [Sat, 22 Dec 2018 08:43:08 +0000 (08:43 +0000)]
llvm-dwarfdump: Remove extraneous space between '(' and 'indexed'

When dumping string or address indexes

llvm-svn: 349997

5 years agollvm-dwarfdump: Print the section name/number for addr_index attributes
David Blaikie [Sat, 22 Dec 2018 08:33:55 +0000 (08:33 +0000)]
llvm-dwarfdump: Print the section name/number for addr_index attributes

(addr attributes coming shortly)

llvm-svn: 349996

5 years agoDebugInfo: Refactor named section dumping into a reusable helper
David Blaikie [Sat, 22 Dec 2018 08:23:10 +0000 (08:23 +0000)]
DebugInfo: Refactor named section dumping into a reusable helper

Currently the section name (& possibly number) is only printed on
addresses in ranges - but no reason it couldn't also be displayed on
other addresses (like low/high PC).

Refactor in that direction by pulling out the section lookup and name
ambiguity dumping logic into a reusable helper.

llvm-svn: 349995

5 years agoAdd support for LLVM profile for NetBSD
Kamil Rytarowski [Sat, 22 Dec 2018 06:56:19 +0000 (06:56 +0000)]
Add support for LLVM profile for NetBSD

Summary:
NetBSD uses typical UNIX interfaces.

All tests pass except instrprof-dlopen-dlclose-gcov.test, as there
is not supported semantics of atexit(3) in dlopen(3)ed+dlclose(3)d
DSO.

NetBSD also ships an older version of LLVM profile (ABI v.2 predating
ABI v.4 in upstream version) inside libc. That copy has been manually
removed during the porting and testing process of the upstream version
to NetBSD. Otherwise there were conflicts between them two.

Reviewers: joerg, vitalybuka, vsk

Subscribers: srhines, fedor.sergeev, llvm-commits, mgorny, #sanitizers

Tags: #sanitizers

Differential Revision: https://reviews.llvm.org/D55997

llvm-svn: 349994

5 years ago[gn build] Embed __TEXT __info_plist section into clang binary on macOS
Nico Weber [Sat, 22 Dec 2018 03:51:10 +0000 (03:51 +0000)]
[gn build] Embed __TEXT __info_plist section into clang binary on macOS

Verified by comparing the output of `otool -P bin/clang` between the GN and the
CMake build.

Differential Revision: https://reviews.llvm.org/D55984

llvm-svn: 349992

5 years ago[gn build] Add build files for clang, clang-offload-bundler, and clang/lib/Headers
Nico Weber [Sat, 22 Dec 2018 03:49:44 +0000 (03:49 +0000)]
[gn build] Add build files for clang, clang-offload-bundler, and clang/lib/Headers

With this, the GN build can build clang!

Differential Revision: https://reviews.llvm.org/D55981

llvm-svn: 349991

5 years agoFix mingw build failures caused by r349839
Tom Stellard [Sat, 22 Dec 2018 03:43:08 +0000 (03:43 +0000)]
Fix mingw build failures caused by r349839

Reviewers: mstorsjo

Subscribers: mgorny, llvm-commits

Differential Revision: https://reviews.llvm.org/D56004

llvm-svn: 349990

5 years ago[gn build] Add build files for llvm-cat, llvm-lto, llvm-lto2, llvm-modextract, llvm...
Nico Weber [Sat, 22 Dec 2018 03:40:35 +0000 (03:40 +0000)]
[gn build] Add build files for llvm-cat, llvm-lto, llvm-lto2, llvm-modextract, llvm-profdata, llvm-symbolizer

These are the llvm/tools needed by check-clang.

Differential Revision: https://reviews.llvm.org/D56018

llvm-svn: 349989

5 years ago[gn build] Add build file for clang/lib/FrontendTool
Nico Weber [Sat, 22 Dec 2018 03:15:56 +0000 (03:15 +0000)]
[gn build] Add build file for clang/lib/FrontendTool

Differential Revision: https://reviews.llvm.org/D55980

llvm-svn: 349988

5 years ago[gn build] Add build file for clang/lib/ARCMigrate
Nico Weber [Sat, 22 Dec 2018 03:15:08 +0000 (03:15 +0000)]
[gn build] Add build file for clang/lib/ARCMigrate

Differential Revision: https://reviews.llvm.org/D55979

llvm-svn: 349987

5 years ago[gn build] Add build files for clang/lib/{ASTMatchers,CrossTU}, clang/lib/StaticAnaly...
Nico Weber [Sat, 22 Dec 2018 03:14:05 +0000 (03:14 +0000)]
[gn build] Add build files for clang/lib/{ASTMatchers,CrossTU}, clang/lib/StaticAnalyzer/{Checkers,Core,Frontend}

The intent is to add the build file for clang/lib/StaticAnalyzer/Frontend; everything else is pulled in by that.

Differential Revision: https://reviews.llvm.org/D55978

llvm-svn: 349986

5 years agoDebugInfo: Remove extra attribute lookup
David Blaikie [Sat, 22 Dec 2018 02:24:13 +0000 (02:24 +0000)]
DebugInfo: Remove extra attribute lookup

llvm-svn: 349985

5 years ago[analyzer] pr38668: Do not attempt to cast loaded integers to floats.
Artem Dergachev [Sat, 22 Dec 2018 02:06:51 +0000 (02:06 +0000)]
[analyzer] pr38668: Do not attempt to cast loaded integers to floats.

This patch is a different approach to landing the reverted r349701.

It is expected to have the same object (memory region) treated as if it has
different types in different program points. The correct behavior for
RegionStore when an object is stored as an object of type T1 but loaded as
an object of type T2 is to store the object as if it has type T1 but cast it
to T2 during load.

Note that the cast here is some sort of a "reinterpret_cast" (even in C). For
instance, if you store an integer and load a float, you won't get your integer
represented as a float; instead, you will get garbage.

Admit that we cannot perform the cast and return an unknown value.

Differential Revision: https://reviews.llvm.org/D55875

rdar://problem/45062567

llvm-svn: 349984

5 years ago[X86] FixupLEAs, reduce number of calls to getOperand and use X86::AddrBaseReg/AddrIn...
Craig Topper [Sat, 22 Dec 2018 01:34:47 +0000 (01:34 +0000)]
[X86] FixupLEAs, reduce number of calls to getOperand and use X86::AddrBaseReg/AddrIndexReg, etc. instead of hardcoded constants.

Makes the code a little more readable.

llvm-svn: 349983

5 years ago[NVPTX] Reduce stack size in NVPTXAsmPrinter::doInitialization().
Justin Lebar [Sat, 22 Dec 2018 01:30:37 +0000 (01:30 +0000)]
[NVPTX] Reduce stack size in NVPTXAsmPrinter::doInitialization().

NVPTXAsmPrinter::doInitialization() was creating an NVPTXSubtarget on
the stack.  This object is huge, about 80kb.  Also it's slow to create.
And it's all redundant; we have one in NVPTXTargetMachine anyway!

llvm-svn: 349982